What's in My LiDAR Odometry Toolbox?

Abstract: With the democratization of 3D LiDAR sensors, precise LiDAR odometries and SLAM are in high demand. New methods regularly appear, proposing solutions ranging from small variations in classical algorithms to radically new paradigms based on deep learning. Yet it is often difficult to compare these methods, notably due to the few datasets on which the methods can be evaluated and compared. Furthermore, their weaknesses are rarely examined, often letting the user discover the hard way whether a method would be appropriate for a use case. In this paper, we review and organize the main 3D LiDAR odometries into distinct categories. We implemented several approaches (geometric based, deep learning based, and hybrid methods) to conduct an in-depth analysis of their strengths and weaknesses on multiple datasets, guiding the reader through the different LiDAR odometries available. Implementation of the methods has been made publicly available at https://github.com/Kitware/pyLiDAR-SLAM.

• The paper presents a comprehensive review comparing classical geometric and deep learning LiDAR odometry methods using diverse datasets.
• It details key methodological steps such as preprocessing, map construction, and energy minimization for robust point cloud registration.
• Hybrid approaches integrating PoseNet initializations with traditional methods are analyzed for improved performance in novel environments.

Overview of "What's in My LiDAR Odometry Toolbox?"

This paper presents a comprehensive review and comparative study of 3D LiDAR odometries, encompassing both classical geometric methods and deep learning approaches. As the use of 3D LiDAR sensors becomes more widespread, the demand for precise LiDAR odometries and SLAM systems increases. However, comparing these methods remains challenging due to the limited datasets and the lack of detailed analysis of their strengths and weaknesses. This study aims to fill that gap by implementing various methods and providing a robust evaluation across multiple datasets.

Classical Geometric SLAM

Classical geometric LiDAR odometries rely on the registration of point clouds against built maps using algorithms like ICP. These methods generally involve motion initialization, point preprocessing, map construction, and neighbor association.

1. Initialization and Preprocessing: Motion initialization can vary widely from simple constant velocity models to more complex strategies incorporating data from cameras, GPS, and IMU sensors. Preprocessing selects candidate points for nearest neighbor pairing and compensates for sensor motion.
2. Map Construction and Neighbor Association: Various strategies construct maps, from voxel grids to surfel maps. Neighbor association methods include projective techniques and tree-based searches, each with trade-offs between precision and computational cost.
3. Energy Minimization: ICP and its variants are essential for aligning point clouds, often requiring careful initialization due to their limited convergence domains.

Deep Learning-Based LiDAR Odometries

Deep learning has introduced new paradigms in LiDAR odometry by using CNN architectures, dubbed PoseNet in this study.

• Supervised and Unsupervised Methods: LO-Net and DeepLO are key methods, with varying approaches to training, using either ground truth supervision or self-supervised techniques that leverage geometric alignment losses.

Hybrid Methods

The paper explores hybrid odometries that combine geometric and learning-based approaches, aiming to leverage the strengths of both. PoseNet can serve as an initial estimate before map-based refinement, potentially improving performance when traditional initialization is challenging.

Datasets and Comparative Analysis

The study employs several datasets—KITTI, Ford Campus, and NCLT—providing diverse scenarios for evaluating the methods. Metrics such as Average Translation Error (ATE) are used to assess performance.

1. KITTI: A mixture of city and highway driving sequences, KITTI serves as a primary dataset. The study reveals that while PoseNet can offer improvements in familiar environments, it struggles with unseen data.
2. Ford Campus and NCLT: These datasets allow further testing of generalization capabilities, highlighting where classical methods still outperform deep learning in novel or complex scenarios.

Key Findings

• PoseNet Limitations: Despite its innovations, PoseNet's generalization remains limited, particularly when encountering novel motions or environments outside its training set.
• Classical vs Deep Learning: Classical methods still prove superior in many real-world scenarios due to their robust, albeit computationally intensive, pipeline.
• Potential of Hybrid Approaches: Hybrid systems show promise by integrating robust initializations via learning-based predictions but require further refinement to become practical.

Implications and Future Directions

The study's findings have implications for both theoretical development and practical applications. It suggests areas where neural networks need improvement in generalization and robustness, particularly for deployment in dynamic real-world environments. The authors intend to extend their modular architecture, pyLiDAR-SLAM, incorporating more advanced techniques to bridge the current performance gap between classical and learning-based odometries.

As the field advances, further refinement of deep learning techniques and their integration with classical methods may yield more adaptive and reliable LiDAR SLAM systems. This work lays the foundation for such innovations by providing a thorough baseline and modular toolkit.